danylo-boiko/stq-ua

Summary

STQ-UA is a large-scale dataset of 100,000 Ukrainian search queries, created via synthetic generation and machine translation. It is designed for training, evaluating, and fine-tuning models on tasks such as query understanding, relevance ranking, and autocorrection.

Details

• Language: Ukrainian (target), English (source for translated portion)
• Total Samples: 100,000 queries
• Granularity: 32 subsegments of 3,125 queries each, balanced by generation method and model
• License: Creative Commons Attribution 4.0 International (CC BY 4.0)
• Authors: Danylo Boiko, Nazar Kohut, Viktoriia Mishkurova, Oleh Basystiuk

Data Fields & Structure

Each entry follows the consistent schema:

Field	Description
query	The unique search query obtained from the model_name.
model_provider	The provider responsible for the model_name (openai, google, cohere, anthropic, or mistral).
model_name	The model used to obtain the query (gpt-4o, gpt-4o-mini, gemini-1.5-flash, gemini-2.0-flash, command-a, command-r-plus, claude-3.5-haiku, or mistral-large).
approach	The approach used to obtain the query (zero-shot, three-shot, or translation).
search_engine	The search engine from which the source_query was taken. If the approach is translation, this field will contain either google or bing.
source_query	The real-world user query taken from search_engine logs.

Citation

If you use this dataset, please cite:

@inproceedings{boiko2026stq-ua,
  author    = {Danylo Boiko and Nazar Kohut and Viktoriia Mishkurova and Oleh Basystiuk},
  title     = {{STQ-UA}: A dataset of synthetic and translated search queries for the Ukrainian language},
  booktitle = {Proceedings of the 2nd International Workshop on Data Analytics (WDA 2026)},
  year      = {2026},
  pages     = {128--137},
  address   = {Kyiv, Ukraine},
  url       = {https://ceur-ws.org/Vol-4155/paper09.pdf},
  publisher = {CEUR-WS.org},
  volume    = {4155}
}